Control automatic skill learning
Decide whether Chetty may turn completed work into personal skills, how often it looks, and how much one review may change.
Each person switches learning on for their own account under Account → Memory with Skill learning. It is off by default and is offered only when the Skills plugin is enabled, the account may create skills, and the change budget is above zero. As an administrator you set the pace and the ceiling for the whole instance.
What a review does
A review reads the completed conversation context of one finished exchange, using the same model and provider account as that conversation, and proposes skills to save.
Under the switch, one muted line reports the last review with its outcome and running totals, and how many successful tool calls remain before the next one, or that no review has run yet. Scheduled, internal, steered and platform-channel turns are never reviewed.
Turn learning on or off
- Open Settings →
<assistant name>AI → Runtime. - Find Maximum skill changes per exchange.
- Set it to
0to switch automatic learning off for everyone, or to a higher value to allow that many changes per review.
0 stops automatic learning only. The manual skill tools keep working, so people can still save and edit skills themselves, and Memory keeps facts separately from skills.
Set the review pace
| Setting | Default | Range | What it does |
|---|---|---|---|
| Minimum tool calls for a skill review | 1 | 1–200 | An exchange with fewer successful tool calls is too simple to hold a procedure, so it is not reviewed or counted. |
| Skill review every N tool calls | 100 | 1–200 | Successful tool calls across an account's eligible exchanges add up to this number, then the next completed exchange is reviewed. |
| Maximum skill changes per exchange | 1 | 0–6 | How many skill writes one review may propose. 0 switches automatic learning off. |
The interval is counted per account and accumulates across exchanges, so a smaller number finds more procedures and costs more model calls. An interval an operator saved earlier stays as saved, even when the default changes.
What a review costs
A review has no access to tools; it only reads and proposes. Its model usage is charged to the account that owns the conversation, and that includes reading the cached conversation context.
A review may be skipped when the exact conversation cannot be safely continued, when no capacity is available, or when the provider account's weekly subscription usage is at least 90% or unknown. A skipped review keeps the count pending and is retried from a later eligible exchange.
Which skills it may change
New learned skills are ordinary personal skills with no approval step, and they behave like any skill the person saved. Turning Show to agent automatically off in the Skills manager keeps one available through /skill:<name> only.
Only a personal skill the agent wrote, still unedited and successfully loaded in the reviewed turn can be improved automatically. Any change a person makes in the manager, including that switch, removes its agent-written status and protects it from later automatic changes. Bundled, shared and manually written skills are never changed.
The Skills plugin validates every proposal;.
Related guides
See Skill learning and memory maintenance for the account side, Skills for the manager, and Memory for durable facts.